CN113109874B - A kind of wave impedance inversion method and neural network system using neural network - Google Patents

Abstract

The present application relates to a wave impedance inversion method and a neural network system using a neural network. The neural network is based on the N-channel seismic data composed of the i-th seismic data and its adjacent multi-channel seismic data, and the i-th initial seismic data. The model data is input, and the i-th wave impedance data is determined as the output, which includes: parallel N feature extraction layers, wherein each feature extraction layer is configured to extract the time series feature of a seismic data input to it; merge The layer is configured to adaptively merge the time series features output by the N feature extraction layers to obtain the spatiotemporal features of the N-channel seismic data; the regression layer is configured to map the spatiotemporal features from the feature domain to the target domain; the output layer is configured as The i-th wave impedance data is determined according to the output of the regression layer and the i-th initial model data. As a result, the wave impedance inversion has higher accuracy, stronger continuity, better noise immunity, and can still maintain a good inversion effect when the initial model is inaccurate.

Description

A wave impedance inversion method using neural network and neural network system

technical field

The present application relates to the technical field of oil and gas exploration, in particular to a wave impedance inversion method using a neural network and a neural network system.

Background technique

Seismic wave impedance inversion is a method of recovering broadband wave impedance data from seismic data with limited frequency bandwidth by comprehensively utilizing existing geological and logging data under the guidance of seismic data. It has been widely used in oil and gas exploration. Reservoir qualitative and quantitative prediction, well pattern deployment, reserve calculation, reservoir dynamic monitoring, etc. in the oil and gas development stage. Because the actual geophysical problems are very complex, and our understanding of them is usually very vague, which makes most of the models established when inverting wave impedance are approximate. The neural network can learn the hidden knowledge from a large amount of existing data, so as to establish a corresponding mathematical model, which makes it very suitable for solving problems where the knowledge background is not clear enough and the model is not accurate enough. Therefore, deep learning algorithms can be applied to wave impedance inversion.

There are many existing wave impedance inversion methods based on deep neural networks, such as wave impedance inversion methods based on convolutional neural networks and wave impedance inversion methods based on closed-loop convolutional neural networks. These networks learn the mapping relationship between seismic data and inversion parameters directly from the training data set. When the underground structure is complex, the relationship between seismic data and inversion parameters is very complicated, and the network generally cannot express this relationship accurately. Although the learning ability of the network can be improved by increasing the depth of the network, due to the limitation of the amount of data, this It increases the risk of network overfitting.

Contents of the invention

In order to solve the above technical problems or at least partly solve the above technical problems, the present application provides an wave impedance inversion method using a neural network and a neural network system.

In a first aspect, the present application provides a wave impedance inversion method using a neural network, the method includes: receiving N channels of seismic data, wherein the N channels of seismic data include the i-th channel of seismic data and its adjacent multiple channels Seismic data; receiving the initial model data of the i-th channel; extracting the time series features of the N channel seismic data by parallel N feature extraction layers of the neural network; adaptively merging the time series features output by the N feature extraction layers by the merging layer of the neural network, Obtain the spatio-temporal features of the seismic data of N channels; the regression layer of the neural network maps the spatio-temporal features from the feature domain to the target domain; the output layer of the neural network determines the wave impedance of the i-th channel according to the output of the regression layer and the initial model data of the i-th channel data.

In some embodiments, before the output layer of the neural network determines the wave impedance data of the i-th track according to the output of the regression layer and the initial model data of the i-th track, it also includes: the initial model data of the i-th track is determined by the correction layer of the neural network Make adaptive adjustments.

In some embodiments, each feature extraction layer includes: a global feature extraction layer configured to extract global features in an input seismic trace; a local feature extraction layer configured to extract local features in an input seismic trace; and/or , the pooling layer is a linearly connected layer with a bias of zero; and/or, the regression layer includes: a set of serial deconvolution blocks configured to upsample the output of the pooling layer; a gated recurrent unit and a fully connected layer, configured to map the upsampled data from the feature domain to the target domain.

In some embodiments, the correction layer is a linearly connected layer and a Tanh activation function.

In a second aspect, the present application provides a neural network system for wave impedance inversion, the neural network system is implemented by one or more computers, and the neural network system is configured to use the i-th seismic data and its adjacent N channels of seismic data composed of multiple channels of seismic data and the i-th channel initial model data are input, and the i-th channel wave impedance data is determined as output. The neural network system includes: parallel N feature extraction layers, wherein, Each feature extraction layer is configured to extract the time-series features of a piece of seismic data input to it; the merging layer is configured to adaptively merge the time-series features output by N feature extraction layers to obtain the spatio-temporal features of N channels of seismic data a regression layer, the regression layer is configured to map the spatiotemporal features from the feature domain to the target domain; an output layer, the output layer is configured to determine the i-th channel wave impedance data according to the output of the regression layer and the i-th channel initial model data.

In some embodiments, the above-mentioned neural network system further includes: a correction layer located before the output layer, and the correction layer is configured to perform adaptive adjustment on the i-th channel initial model data.

In some embodiments, each feature extraction layer includes: a global feature extraction layer configured to extract global features in input seismic traces; a local feature extraction layer configured to extract Local features in the input seismic traces; and/or, the merging layer is a linearly connected layer with a bias of zero; and/or, the regression layer includes: a set of serial deconvolution blocks configured to perform Upsampling; gated recurrent units and fully connected layers configured to map the upsampled data from the feature domain to the target domain.

In some embodiments, the correction layer is a linearly connected layer and a Tanh activation function.

In a third aspect, the present application provides a method for training a neural network for wave impedance inversion, including: receiving first input data and second input data, wherein the first input data includes: the ith channel seismic data and N channels of seismic data composed of adjacent multiple channels of seismic data, and the i-th channel initial model data; the second input data includes: seismic data at the well location and wave impedance data at the well location; the inversion neural network is based on The first input data determines the i-th channel wave impedance data, the first synthetic seismic data is determined by the forward modeling neural network based on the i-th channel wave impedance data output by the inversion neural network, and the synthetic first seismic data is determined by the forward modeling neural network according to the input well position The i-th channel wave impedance data determines the synthesized second seismic data; determines the first mean square error between the input i-th channel seismic data and the synthesized first seismic data, and determines the i-th channel at the well location by the inversion neural network The second mean square error between the output wave impedance data and the input wave impedance data, and the third mean square error between the second seismic data synthesized in the i-th track at the well location and the input seismic data; using the first The first mean square error is used to update the parameters of the inversion neural network and the forward modeling neural network, the second mean square error is used to update the parameters of the inversion neural network, and the third mean square error is used to update the parameters of the forward modeling neural network.

In some embodiments, the above-mentioned inversion neural network includes: N feature extraction layers juxtaposed, wherein each feature extraction layer is configured to extract the time series features of a piece of seismic data input to it; the merging layer, the merging The layer is configured to adaptively merge the time series features output by N feature extraction layers to obtain the temporal and spatial characteristics of N channels of seismic data; the regression layer is configured to map the temporal and spatial characteristics from the feature domain to the target domain; the output layer, the The output layer is configured to determine the i-th wave impedance data according to the output of the regression layer and the i-th initial model data; and/or, the above-mentioned forward neural network includes: a set of serial one-dimensional convolutional layers and activation functions .

Compared with the prior art, the above-mentioned technical solution provided by the embodiment of the present application has the following advantages: the method provided by the embodiment of the present application has higher accuracy of wave impedance inversion, stronger continuity, and better noise resistance. And it can still maintain a good inversion effect even when the initial model is not accurate.

Description of drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and together with the description serve to explain the principles of the invention.

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, In other words, other drawings can also be obtained from these drawings without paying creative labor.

Fig. 1 is the structural block diagram of an embodiment of the wave impedance inversion system provided by the embodiment of the present application;

Fig. 2 is a flow chart of an embodiment of an wave impedance inversion method using a neural network provided in an embodiment of the present application;

FIG. 3 is a schematic diagram of an embodiment of a system for training a neural network provided in an embodiment of the present application;

FIG. 4 is a flow chart of an embodiment of a method for training a neural network for wave impedance inversion provided in an embodiment of the present application;

Fig. 5 is the comparison diagram of the wave impedance inversion result and the real wave impedance on the Marmousi2 model by applying the inversion method proposed in the embodiment of the present application;

Fig. 6 is a wave impedance inversion result graph of actual data; and

FIG. 7 is a schematic hardware diagram of an implementation manner of a computer device provided in an embodiment of the present application.

Detailed ways

It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

In the following description, use of suffixes such as 'module', 'part' or 'unit' for denoting elements is only for facilitating description of the present invention and has no specific meaning by itself. Therefore, 'module', 'part' or 'unit' may be used in combination.

Wave Impedance Inversion Using Neural Networks

The embodiment of the present application provides a neural network for wave impedance inversion. Since the logging information and geological structure information are generally used comprehensively when establishing the initial model, the initial model can reflect the real structure of the underground to a certain extent. And there is a certain correlation in the underground structure, which is related to the distance, the closer the distance, the stronger the correlation, and vice versa, the weaker the correlation. Since the depth of the seismic section is usually expressed as time depth, the spatial correlation of the subterranean structure on the seismic section is the time correlation of the seismic traces in the vertical direction and the spatial correlation between the central trace and adjacent traces in the horizontal direction. Therefore, in the embodiment of the present application, formula (1) is used to represent the relationship between the underground real wave impedance and the initial model and seismic data:

为初始模型中第i道上t时刻处的波阻抗值；{x_i-k,t,…,x_i-1,t,x_i,t,…,x_i+k,t}为第i道与相邻的2k道地震数据在t时刻的振幅序列，应当理解，在某些实施例中不限于与第i道相邻的2k道地震数据，任意多道也是可行的；g(·)为初始模型修正函数，它用于修正初始模型以减少初始模型与地下真实的波阻抗之间的差异，应当理解，在某些实施例中可不修正初始模型，本申请实施例对此不做限定；f(·)为残差函数，它用于表示真实波阻抗值与初始模型中的波阻抗值之间的误差值。In formula (1), m _i,t is the real wave impedance value at time t on the i-th track;

is the wave impedance value at time t on the i-th track in the initial model; {xi _ik,t ,…, _xi-1,t , _xi,t ,…, _xi+k,t } is the The amplitude sequence of the adjacent 2k channels of seismic data at time t, it should be understood that in some embodiments, it is not limited to the 2k channels of seismic data adjacent to the i-th channel, and any number of channels is also feasible; g(·) is the initial model A correction function, which is used to correct the initial model to reduce the difference between the initial model and the real underground wave impedance. It should be understood that in some embodiments, the initial model may not be corrected, which is not limited in the embodiment of the present application; f( ) is the residual function, which is used to represent the error value between the real wave impedance value and the wave impedance value in the initial model.

Based on the relationship between the real wave impedance and the initial model and seismic data in formula (1), the present application constructs a neural network model to learn this mapping relationship.

Figure 1 is a structural block diagram of an embodiment of the wave impedance inversion system provided by the embodiment of the present application. As shown in Figure 1, the wave impedance inversion system includes: neural network input 110, inversion neural network system 120 and neural network output 130.

The neural network input 110 includes: N channels of seismic data, the N channels of seismic data including the i-th channel of seismic data and multiple channels of adjacent seismic data; and the i-th channel of initial model data. As shown in Fig. 1, the N channels of seismic data are the i-th channel of seismic data and 2k channels of seismic data adjacent to it, wherein there are k channels of seismic data on both sides. It should be understood that this embodiment of the present application does not limit this, and multiple channels of seismic data adjacent to the i-th channel of seismic data are all feasible. The neural network output 130 is the i-th channel wave impedance data, which is the predicted value of the i-th channel wave impedance.

The inversion neural network system 120 , as shown in FIG. 1 , the inversion neural network system 120 includes: N feature extraction layers 121 , merging layer 122 , regression layer 123 , correction layer 124 and output layer 125 arranged in parallel.

Each feature extraction layer 121 is configured to extract time series features of a piece of seismic data input thereto. Each feature extraction layer 121 processes one seismic data, and N feature extraction layers 121 parallel to process N seismic data in parallel, as shown in Figure 1, is 2k+1 seismic data, and the neural network system 120 includes 2k+1 feature extraction layer. The timing feature represents the correlation of seismic traces in time.

In some embodiments, as shown in FIG. 1 , each feature extraction layer 121 includes a global feature extraction layer 1211 and a local feature extraction layer 1212, wherein the global feature extraction layer 1211 is configured to extract global features of the input seismic trace, The local feature extraction layer 1212 is configured to extract local features in the input seismic traces. As an example, the global feature extraction layer 1211 is composed of a set of serial bidirectional gated recurrent units (GRU); the local feature extraction layer 1212 includes: a set of parallel convolution blocks with different expansion coefficients, configured to extract different scales in the seismic trace The local features of ; fully connected layers and convolutional blocks are configured to combine the extracted local features. It should be understood that in the embodiment of the present application, other structures capable of extracting temporal correlation of seismic traces may also be conceived, which is not limited in the embodiment of the present application.

Referring to FIG. 1 , each feature extraction layer 121 outputs time-series features corresponding to its input seismic traces, and N feature extraction layers 121 output N time-series features corresponding to N seismic traces. The merging layer 122 is configured to adaptively merge the time series features output by the N feature extraction layers 121 to obtain the spatiotemporal features of N channels of seismic data. Spatio-temporal features not only represent the correlation of seismic traces in time, but also represent the spatial correlation between seismic traces and adjacent seismic traces. In some embodiments, the merging layer 122 is a linearly connected layer with a bias of zero, but this is not limited in this embodiment of the present application.

After the neural network input 110 extracts temporal features through N feature extraction layers 121 , it outputs spatiotemporal features via merging layer 122 , and the spatiotemporal features are used as the input of regression layer 123 . The regression layer 123 is configured to map spatio-temporal features from the feature domain to the target domain. In some embodiments, the regression layer 123 includes: a set of serial deconvolution blocks configured to upsample the output of the merging layer 122 to have a preset sampling rate; a GRU and a fully connected layer, It is configured to map the upsampled data from the feature domain to the target domain.

The parallel N feature extraction layers 121 , merging layer 122 and regression layer 123 complete the residual modeling of the time-space sequence of the network and realize the residual function f(·) in formula (1). The output of the regression layer 123 and the i-th initial model data are jointly input to the output layer 125 . The output layer 125 is configured to determine the i-th channel wave impedance data according to the output of the regression layer 123 and the i-th channel initial model data. In some embodiments, the output layer 125 adds the i-th channel of initial model data to the output of the regression layer 123 to obtain the i-th channel of wave impedance data, but the embodiment of the present application is not limited thereto.

In some embodiments, as shown in FIG. 1 , before the i-th channel initial model data is input to the output layer, it is also corrected by the correction layer 124 located before the output layer 125 . The correction layer 124 is configured to perform adaptive adjustment on the i-th channel initial model data to realize the correction function g(·) in formula (1). In some embodiments, the revision layer 124 includes a linear connection layer and a Tanh activation function.

As a preferred example, the global feature extraction layer 1211 is composed of a set of serial bidirectional GRUs to extract global features in the input seismic trace. The local feature extraction layer 1212 first uses a set of parallel convolutional blocks with different expansion coefficients to extract local features of different scales in the input seismic trace, and then uses the fully connected layer and convolutional blocks to combine these local features. The global feature extraction layer 1211 and the local feature extraction layer 1212 constitute the time series feature extraction module of the network, which can extract the time series features of the input seismic trace. Apply 2k+1 time-series feature extraction modules to extract the time-series features of the i-th channel seismic data and its adjacent 2k channels of seismic data, and then use the merging layer 122 to adaptively merge the extracted features, thereby obtaining the temporal-spatial features of the seismic data feature, where the pooling layer 122 consists of a linearly connected layer with a bias of zero. The regression layer 123 first uses a set of serial deconvolution blocks to upsample the output of the merging layer so that it has a preset sampling rate (for example, the same sampling rate as the label data during training), and then uses a GRU and A fully connected layer maps the upsampled data from the feature domain to the target domain. The regression layer 123 and the spatiotemporal feature extraction module constitute the spatiotemporal sequence residual modeling module of the network. The correction layer 124 uses the linear connection layer and the Tanh activation function to adaptively adjust the initial model data of the i-th channel. Finally, the output of the time-space sequence residual modeling module is added to the output of the correction layer 124 to obtain the predicted wave impedance data.

An embodiment of the present application provides a wave impedance inversion method using a neural network. As shown in FIG. 2 , the wave impedance inversion method includes steps S202 to S212.

Step S202, receiving N channels of seismic data, wherein the N channels of seismic data include the i-th channel of seismic data and its adjacent multiple channels of seismic data;

Step S204, receiving the i-th track initial model data.

In step S206, the time-series features of the N channels of seismic data are extracted by the parallel N feature extraction layers of the neural network.

In step S208, the merging layer of the neural network adaptively merges the time series features output by the N feature extraction layers to obtain the temporal and spatial features of the N channels of seismic data.

In step S210, the regression layer of the neural network maps the spatio-temporal features from the feature domain to the target domain.

In step S212, the output layer of the neural network determines the i-th wave impedance data according to the output of the regression layer and the i-th initial model data.

It should be understood that although the sequence numbers of the steps are marked in FIG. 2 , this is not a limitation on the execution sequence of the steps. For example, step S204 may be performed at any step before step S212.

In some embodiments, before the above step S212, it also includes: adaptively adjusting the initial model data of the i-th track by the correction layer of the neural network. In some embodiments, the correction layer is a linearly connected layer and a Tanh activation function.

In some embodiments, the above step S206 includes: extracting global features in the input seismic trace by the global feature extraction layer of the feature extraction layer, and extracting local features in the input seismic trace by the local feature extraction layer of the feature extraction layer.

In some embodiments, the merging layer used in the above step S208 is a linearly connected layer with a bias of zero.

In some embodiments, in the above step S210, the output of the combining layer is first up-sampled by a set of serial deconvolution blocks of the regression layer, and then the up-sampling is performed by the gated recurrent unit and the fully connected layer of the regression layer. The sampled data is mapped from the feature domain to the target domain.

In some embodiments, in step S212, the output layer of the neural network adds the output of the regression layer and the i-th channel of initial model data to obtain the i-th channel of wave impedance data.

In the embodiment of the present application, the neural network may refer to the network structure shown in FIG. 1 , which will not be described in detail here.

As a preferred example, a set of serial bidirectional GRUs extract the global features of the input seismic traces, use a set of parallel convolution blocks with different expansion coefficients to extract local features of different scales in the input seismic traces, and then use the fully connected layer and Convolution blocks are used to combine these local features, thereby extracting the time series features of the input seismic traces. Extract the time-series features of the i-th seismic data and its adjacent 2k seismic data, and then use the merge layer to adaptively merge the extracted features to obtain the spatial-temporal features of the seismic data, where the merge layer is offset by one to zero of linearly connected layers. Use a set of serial deconvolution blocks to upsample the output of the pooling layer to a preset sampling rate (e.g., the same sampling rate as the label data during training), and then use a GRU and a fully connected layer Map the upsampled data from feature domain to target domain. The initial model data of the i-th channel is adaptively adjusted by using a linear connection layer and a Tanh activation function. Finally, the output of the regression layer is added to the output of the correction layer to obtain the predicted wave impedance data.

Neural Network Training

Considering the lack of labeled data in actual exploration, the neural network is trained by semi-supervised learning in the embodiment of the present application. Fig. 3 is a schematic diagram of an embodiment of a system for training a neural network provided in an embodiment of the present application. As shown in Fig. 3, the system includes: a first input 310, a second input 320, an inversion neural network 330 and a forward neural network 340.

As shown in FIG. 3 , the first input 310 is used to input the first input data, wherein the first input data includes N channels of seismic data composed of the i-th channel of seismic data and multiple channels of seismic data adjacent to it, and the i-th channel of seismic data Figure 3 shows the seismic data of the i-th channel and the k channels of seismic data on both sides of the i-th channel, a total of 2k+1 seismic data.

As shown in FIG. 3 , the second input 320 is used to input second input data, wherein the second input data includes acoustic impedance data 321 at the well location and seismic data 322 at the well location. The acoustic impedance data 321 at the well location is the actual acoustic impedance value (also called the input acoustic impedance data), and the seismic data 322 at the well location is the actual seismic data (also called the input seismic data).

As shown in FIG. 3 , the inversion neural network 330 is configured to determine the i-th channel of wave impedance data 331 corresponding to the i-th channel of seismic data according to the first input data. The forward modeling neural network 340 is configured to determine the synthesized i-th channel seismic data 341a according to the output of the inversion neural network 330 (i-th channel wave impedance data 331), and determine the synthesized i-th channel seismic data 341a according to the input i-th channel wave impedance data 321 i channel seismic data 341b.

As shown in FIG. 3 , the system further includes: a first error determination module 350 configured to determine a first mean square error (l _seismic ) between the input i-th channel seismic data and the synthesized i-th channel seismic data 341a The second error determination module 360 is configured to determine the second mean square error (l _well ) between the wave impedance data 331 and the wave impedance data 321 of the i track at the well position; the third error determination module 370 is It is configured to determine a third mean square error (l' _well ) between the seismic data 341b and the seismic data 322 at the well location for the ith track.

As shown in FIG. 3 , the internal parameters of the inversion neural network 330 are updated using the first mean square error and the second mean square error, and the internal parameters of the forward modeling neural network 340 are updated using the first mean square error and the third mean square error.

In the embodiment of the present application, the inversion neural network 330 may refer to the network structure shown in FIG. 1 , which will not be described in detail here. The forward neural network 340 uses a set of serial one-dimensional convolution layers and activation functions to simulate the forward process of the convolution model.

An embodiment of the present application provides a method for training a neural network for wave impedance inversion. As shown in FIG. 4 , the method includes steps S402 to S416.

Step S402, receiving first input data and second input data.

Wherein, the first input data includes N-channel seismic data formed by the i-th channel seismic data and its adjacent multi-channel seismic data, and the i-th channel initial model data; the second input data includes: seismic data at the well position and wave impedance data at the well location.

In step S404, the i-th channel wave impedance data is determined by the inversion neural network according to the first input data.

In step S406, the synthetic first seismic data is determined by the forward neural network based on the i-th channel wave impedance data output by the inverse neural network.

In step S408, the synthetic second seismic data is determined by the forward modeling neural network according to the input i-th channel wave impedance data at the well location.

Step S410, determining a first mean square error between the input i-th track of seismic data and the synthesized first seismic data.

Step S412, determining the second mean square error between the wave impedance data output by the inversion neural network and the input wave impedance data at the i-th channel at the well location.

Step S414, determining the third mean square error between the i-th channel synthesized second seismic data and the input seismic data at the well location.

Step S416, updating the parameters of the inversion neural network and the forward neural network.

Wherein, the first mean square error is used to update the parameters of the inversion neural network and the forward modeling neural network, the second mean square error is used to update the parameters of the inversion neural network, and the third mean square error is used to update the parameters of the forward modeling neural network.

In some embodiments, the above-mentioned inversion neural network includes: N feature extraction layers juxtaposed, wherein each feature extraction layer is configured to extract the time series features of a piece of seismic data input to it; the merging layer, the merging The layer is configured to adaptively merge the time series features output by N feature extraction layers to obtain the temporal and spatial characteristics of N channels of seismic data; the regression layer is configured to map the temporal and spatial characteristics from the feature domain to the target domain; the output layer, the The output layer is configured to determine the i-th channel wave impedance data according to the output of the regression layer and the i-th channel initial model data.

In some embodiments, the inversion neural network further includes a correction layer configured to perform adaptive adjustment on the i-th channel initial model data. The output layer is configured to determine the i-th channel wave impedance data according to the output of the regression layer and the corrected i-th channel initial model data (the output of the correction layer).

In the embodiment of the present application, the inversion neural network may refer to the network structure shown in FIG. 1 , which will not be described in detail here.

In some embodiments, the above-mentioned forward neural network includes a set of serial one-dimensional convolution layers and activation functions to simulate the forward process of the convolution model.

As a preferred example, the global feature extraction layer of the inversion neural network is composed of a group of serial bidirectional GRUs (Gated Recurrent Units) for extracting global features in the input seismic trace. The local feature extraction layer of the inversion neural network first uses a set of parallel convolutional blocks with different expansion coefficients to extract local features of different scales in the input seismic trace, and then uses the fully connected layer and convolutional blocks to combine these local features. The two constitute the time series feature extraction module of the inversion neural network, which can extract the time series features of the input seismic trace. Apply 2k+1 time-series feature extraction modules to extract the time-series features of the i-th channel seismic data and its adjacent 2k channel seismic data, and then use the merging layer of the inversion neural network to adaptively merge the extracted features to obtain Spatiotemporal characterization of seismic data, where the pooled layer consists of a linearly connected layer with a bias of zero. The regression layer of the inversion neural network first uses a set of serial deconvolution blocks to upsample the output of the pooling layer to have the same sampling rate as the label data, and then uses a GRU and a fully connected layer to upsample The final data is mapped from the feature domain to the target domain. It and the spatiotemporal feature extraction module constitute the spatiotemporal sequence residual modeling module of the inversion neural network. The correction layer of the inversion neural network adopts the linear connection layer and the Tanh activation function to adaptively adjust the i-th initial model data. Finally, adding the output of the time-space sequence residual modeling module and the output of the correction layer is the wave impedance data predicted by the network. The forward neural network consists of a set of serial one-dimensional convolutional layers and activation functions to simulate the forward process of the convolution model. During training, the update of the internal parameters of the inversion neural network and the forward neural network is affected by the combined effects of the following two processes:

1) The inversion neural network predicts the wave impedance data of the i-th channel according to the input seismic data of the i-th channel and its adjacent 2k seismic data and the initial model data of the i-th channel, and then the predicted wave impedance of the i-th channel The data is fed into a forward neural network to obtain synthetic seismic data. The internal parameters of the inversion neural network and the forward modeling neural network are updated by calculating the mean square error l _seismic between the synthetic i-th channel seismic data and the input i-th channel seismic data. All seismic traces participate in this process.

2) The inversion neural network predicts the wave impedance at the well location according to the input seismic data at the well location, the 2k channel seismic data adjacent to the well location, and the initial model data at the well location, and calculates the predicted well location. The mean square error l _well between the wave impedance data and the actual logging wave impedance data is used to update the parameters in the inversion neural network. The forward modeling neural network predicts the corresponding seismic records according to the wave impedance in the actual logging, and updates the forward modeling neural network by calculating the mean square error l' _well between the seismic data at the well location and the seismic data predicted by the forward modeling neural network. parameters. Only the seismic traces at the well site are involved in this process.

In this preferred example, the loss functions shown in formulas (2), (3) are constructed and the Adam optimizer is used to update the learnable parameters in the inversion neural network and the forward neural network:

是正演模型输出的合成的第i道地震数据，

是井中的波阻抗数据，

是反演神经网络所预测的井位处的波阻抗数据，

是井位处的地震数据，

是正演神经网络所预测的井位处的地震数据，L(·)是均方误差函数，其定义如公式(4)所示，α、β是权重系数，它们的值可以根据地震数据与测井数据的质量进行调整。Where x _i,t is the input i-th seismic data,

is the synthetic i-th channel seismic data output by the forward modeling model,

is the wave impedance data in the well,

is the wave impedance data at the well location predicted by the inversion neural network,

is the seismic data at the well location,

is the seismic data at the well location predicted by the forward modeling neural network, L( ) is the mean square error function, and its definition is shown in formula (4), α and β are weight coefficients, and their values can be calculated according to the seismic data and measured The quality of the well data is adjusted.

Fig. 5 is the comparison of the wave impedance inversion result and the real wave impedance on the Marmousi2 model using the inversion method proposed in the embodiment of the present application, wherein, (a) real wave impedance, (b) wave impedance inversion result, ( c) Error between true wave impedance and predicted wave impedance. Fig. 6 is the wave impedance inversion result of actual data. It can be seen that the inversion result based on the method proposed in the embodiment of the present application has a small error with the real wave impedance on the model data. In the actual data, the inversion result has a relatively high resolution, and at the test well , the inversion results are basically consistent with the change trend of the wave impedance of the test well.

Through the embodiment of this application, a space-time sequence residual modeling network is constructed. The network takes the initial model as the initial value, continuously corrects the initial model during the learning process, and learns the residual between the corrected initial model and the inversion parameters. Poor, and fully consider the spatiotemporal characteristics of the data in the residual learning process. Considering the lack of labeled data in actual exploration, the embodiment of the present application adopts a semi-supervised learning method to train the network model. The trained network predicts wave impedance data based on initial model data and seismic data.

In comparison, the semi-supervised learning seismic inversion in related technologies can well mine the information in logging data and seismic data, but the utilization rate of the initial model is poor, and the initial model is generally only used as a label To constrain the inversion of the network, which makes the network unable to make good use of the structural information and rich low-frequency information contained in the initial model.

In addition, there is a certain correlation in the underground structure in space, which is reflected in the spatial correlation in the horizontal direction and the temporal correlation in the vertical direction on the seismic section. The deep learning inversion neural network in the related technology fully exploits the vertical time correlation in seismic data, but does not consider the horizontal spatial correlation of seismic data, which leads to the inconsistency of the network prediction results. Horizontal continuity is poor.

The embodiment of the present application also provides a computer device. The computer device 20 in this embodiment at least includes but is not limited to: a memory 21 and a processor 22 that can be communicated with each other through a system bus, as shown in FIG. 7 . It should be noted that FIG. 7 only shows computer device 20 having components 21-22, but it should be understood that implementing all of the illustrated components is not required and that more or fewer components may instead be implemented.

In this embodiment, the memory 21 (that is, a readable storage medium) includes a flash memory, a hard disk, a multimedia card, a card-type memory (for example, SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), Read Only Memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), Programmable Read Only Memory (PROM), Magnetic Memory, Magnetic Disk, Optical Disk, etc. In some embodiments, the memory 21 may be an internal storage unit of the computer device 20 , such as a hard disk or memory of the computer device 20 . In some other embodiments, the memory 21 can also be an external storage device of the computer device 20, such as a plug-in hard disk equipped on the computer device 20, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, flash memory card (Flash Card), etc. Of course, the storage 21 may also include both the internal storage unit of the computer device 20 and its external storage device. In this embodiment, the memory 21 is generally used to store the operating system installed in the computer device 20 and various application software, such as the program code of the wave impedance inversion method and the like. In addition, the memory 21 can also be used to temporarily store various types of data that have been output or will be output.

The processor 22 may be a central processing unit (Central Processing Unit, CPU), a controller, a microcontroller, a microprocessor, or other data processing chips in some embodiments. The processor 22 is generally used to control the overall operation of the computer device 20 . In this embodiment, the processor 22 is configured to run the program codes stored in the memory 21 or process data, for example, the program codes of the wave impedance inversion method, so as to realize the wave impedance inversion method.

This embodiment also provides a computer-readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), only Read memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, server, App application store, etc., on which computer programs are stored, The corresponding functions are realized when the program is executed by the processor. The computer-readable storage medium in this embodiment is used to store the wave impedance inversion program, and when executed by the processor, implements the steps of the wave impedance inversion method.

It should be noted that, in this document, the term "comprising", "comprising" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article or apparatus comprising a set of elements includes not only those elements, It also includes other elements not expressly listed, or elements inherent in the process, method, article, or device. Without further limitations, an element defined by the phrase "comprising a ..." does not preclude the presence of additional identical elements in the process, method, article, or apparatus comprising that element.

The serial numbers of the above embodiments of the present invention are for description only, and do not represent the advantages and disadvantages of the embodiments.

Through the description of the above embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus a necessary general-purpose hardware platform, and of course also by hardware, but in many cases the former is better implementation. Based on this understanding, the technical solution of the present invention can be embodied in the form of a software product in essence or the part that contributes to the prior art, and the computer software product is stored in a storage medium (such as ROM/RAM, disk, CD) contains several instructions to make a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) execute the methods described in various embodiments of the present invention.

Embodiments of the present invention have been described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementations, and the above-mentioned specific implementations are only illustrative, rather than restrictive, and those of ordinary skill in the art will Under the enlightenment of the present invention, many forms can also be made without departing from the gist of the present invention and the protection scope of the claims, and these all belong to the protection of the present invention.

Claims (7)

1. A wave impedance inversion method using a neural network, characterized in that, the wave impedance inversion method comprises: Receive N channels of seismic data, wherein the N channels of seismic data include the i-th channel of seismic data and multiple channels of seismic data adjacent to it; Receive the i-th channel initial model data; Extracting the time-series features of the N channels of seismic data by the parallel N feature extraction layers of the neural network; The time-series features output by the N feature extraction layers are adaptively merged by the merging layer of the neural network to obtain the temporal and spatial characteristics of the N channels of seismic data. The temporal and spatial characteristics not only represent the temporal correlation of the seismic channels, but also Indicates the spatial correlation between a seismic trace and adjacent seismic traces; Mapping the spatio-temporal features from the feature domain to the target domain by the regression layer of the neural network; Determining the i-th channel wave impedance data according to the output of the regression layer and the i-th channel initial model data by the output layer of the neural network; Before determining the wave impedance data of the i-th channel according to the output of the regression layer and the initial model data of the i-th channel by the output layer of the neural network, it also includes: correcting the i-th channel by the correction layer of the neural network The initial model data is adaptively adjusted, the initial model is used as the initial value, the initial model is continuously corrected during the learning process and the residual between the corrected initial model and the inversion parameters is learned, and the residual error is fully considered in the residual learning process The spatiotemporal characteristics of the data. 2. wave impedance inversion method according to claim 1, is characterized in that, Each feature extraction layer includes: a global feature extraction layer configured to extract global features in an input seismic trace; a local feature extraction layer configured to extract local features in an input seismic trace; and/or The pooling layer is a linearly connected layer with a bias of zero; and/or The regression layer includes: a set of serial deconvolution blocks configured to upsample the output of the pooling layer; a gated recurrent unit and a fully connected layer configured to map the upsampled data from the feature domain to target domain. 3. The wave impedance inversion method according to claim 1, wherein the correction layer is a linear connection layer and a Tanh activation function. 4. A neural network system for wave impedance inversion, characterized in that, the neural network system is implemented by one or more computers, and the neural network system is configured to use the i-th channel seismic data and its adjacent The N-channel seismic data formed by the multi-channel seismic data and the i-th channel initial model data are input, and the i-th channel wave impedance data is determined as an output, wherein the neural network system includes: N feature extraction layers juxtaposed, wherein each feature extraction layer is configured to extract time series features of a piece of seismic data input thereto; A merging layer, the merging layer is configured to adaptively merge the time-series features output by the N feature extraction layers to obtain the temporal and spatial features of the N channels of seismic data, and the temporal and spatial features not only represent the temporal correlation of the seismic channels , also represents the spatial correlation between a seismic trace and adjacent seismic traces; a regression layer configured to map the spatiotemporal features from a feature domain to a target domain; an output layer, the output layer is configured to determine the i-th channel wave impedance data according to the output of the regression layer and the i-th channel initial model data; The neural network system also includes: a correction layer located before the output layer, the correction layer is configured to adaptively adjust the i-th initial model data, with the initial model as the initial value, during the learning process Continuously modify the initial model and learn the residual between the modified initial model and the inversion parameters, and fully consider the spatiotemporal characteristics of the data in the process of residual learning. 5. neural network system according to claim 4, is characterized in that, Each feature extraction layer includes: a global feature extraction layer configured to extract global features in the input seismic trace; a local feature extraction layer configured to extract local features in the input seismic trace features; and/or The pooling layer is a linearly connected layer with a bias of zero; and/or The regression layer includes: a set of serial deconvolution blocks configured to upsample the output of the pooling layer; a gated recurrent unit and a fully connected layer configured to map the upsampled data from the feature domain to target domain. 6. The neural network system according to claim 4, wherein the correction layer is a linear connection layer and a Tanh activation function. 7. A method for training a neural network for wave impedance inversion, comprising: Receive first input data and second input data, wherein the first input data includes: N channels of seismic data composed of the i-th channel of seismic data and multiple channels of seismic data adjacent to it, and the i-th channel of initial model data ; The second input data includes: seismic data at the well location and wave impedance data at the well location; Determining the i-th channel wave impedance data by the inversion neural network according to the first input data, determining the synthesized first seismic data by the forward modeling neural network according to the i-th channel wave impedance data output by the inversion neural network, and by The forward modeling neural network determines the synthetic second seismic data according to the i-th channel wave impedance data at the input well location; Determine the first mean square error between the input i-th channel seismic data and the first synthetic seismic data, determine the i-th channel wave impedance data output by the inversion neural network at the well position and the input wave impedance a second mean square error between the data, and a third mean square error between the synthetic second seismic data of the i-th track at the well location and the seismic data at the well location; Using the first mean square error to update the parameters of the inversion neural network and the forward neural network, using the second mean square error to update the parameters of the inversion neural network, and using the third mean square error square error updates the parameters of the forward modeling neural network; The inversion neural network includes: N feature extraction layers juxtaposed, wherein each feature extraction layer is configured to extract the time series features of a piece of seismic data input to it; a merging layer, the merging layer is configured to automatically Adapting to merging the time series features output by the N feature extraction layers to obtain the temporal and spatial characteristics of the N seismic data, the temporal and spatial characteristics not only represent the temporal correlation of seismic traces, but also represent the relationship between seismic traces and adjacent seismic traces. The spatial correlation between; the regression layer, the regression layer is configured to map the spatio-temporal features from the feature domain to the target domain; the output layer, the output layer is configured to be based on the output of the regression layer and the first The i-th channel of initial model data determines the i-th channel of wave impedance data; a correction layer located before the output layer, the correction layer is configured to perform adaptive adjustment on the i-th channel of initial model data, using the initial model as an initial value , continuously modifying the initial model during the learning process and learning the residual between the modified initial model and the inversion parameters, and fully considering the spatio-temporal characteristics of the data during the residual learning process; and/or The forward neural network includes: a set of serial one-dimensional convolutional layers and activation functions.

Images